IBM and NASA Launch a Lunar Foundation Model Trained on Decades of Space Data

Morning, {{first name | folks}}! AI is mapping the Moon, hackers are using AI to rewrite malware when it gets detected, Codex is helping search for new antibiotics, Cursor is coordinating thousands of coding agents, and Gemini just landed on Windows. Let's get into it.

IBM and NASA open-sourced a new lunar foundation model trained on decades of data from NASA and Japanese missions. The model can work across different sensors, resolutions, and viewing angles, with NASA using it to map small craters, study volcanic features, and identify areas near the lunar poles that could contain ice. In testing, it reduced ice-detection error by 22% compared with a leading transformer model.

Lunar missions generate huge amounts of data from different instruments, but researchers still have to bring those datasets together before they can use them. The new model gives them one system for working across that data and narrowing down where to investigate. That could make existing lunar data much more useful before the next mission even leaves Earth.

Anthropic's third threat intelligence report documents real AI misuse disrupted between December 2025 and August 2026, and the most striking case involves a suspected Russian state-linked group that used Claude to monitor its own malware. The moment security tools flagged it, AI agents rewrote and redeployed the code automatically, cycling until nothing was detected. Anthropic's own framing, this "inverts the cost back onto defenders," the tool that used to slow attackers down, just got outpaced by them.

Elsewhere in the report, one criminal group turned a single vendor breach into access across roughly 200 downstream companies in about 34 hours, and a lone hacktivist built a functioning mass surveillance platform entirely by themselves. One genuinely reassuring detail buried in all of it, none of the disrupted cases involved Anthropic's most capable models, the safeguards on those actually held.

OpenAI is using Codex with researchers at the de la Fuente Lab to search for antimicrobial molecules that could fight drug-resistant bacteria. Codex helped analyze biological data, write and run code, investigate potential candidates, and work through the research needed to narrow down what was worth testing.

The candidates still have to prove themselves in the lab, but the researchers say Codex can take on large parts of the search that would otherwise take much longer. Instead of just helping scientists read research, it’s now helping them decide what to investigate next.

Cursor introduced Projects, a new way to give AI agents larger software projects instead of individual coding tasks. A coordinator agent can break work into smaller jobs and send them to thousands of subagents running in parallel while keeping shared context across the project. Agents can run in the cloud, work from local environments, and coordinate through the same project.

Projects can also respond to events outside the editor, including Slack messages, pull requests, scheduled tasks, and CI failures. That means an agent can keep working on a project without someone having to start a new prompt every time. Cursor is pushing its coding agents further into the day-to-day software workflow, where work can continue in the background rather than stopping when the developer closes the editor.

Google launched a dedicated Gemini app for Windows, giving users access to Gemini directly from the desktop with Alt + Space. The app can work with files and apps on the PC, connect to Gmail and Google Drive, and run tasks through Gemini’s agent features. Google is making it available on Windows 10 and 11.

The bigger change is where Gemini shows up. Instead of asking people to leave what they’re doing and open an AI website, Google can put Gemini alongside the apps they already use for work. That gives it a much easier route into everyday tasks, files, and information. AI becomes part of the desktop workflow rather than another destination users have to remember to open.